Towards Identifiable Unsupervised Domain Translation: A Diversified Distribution Matching Approach